3 * THE REGENTS OF THE UNIVERSITY OF MICHIGAN
6 * Permission is granted to use, copy, create derivative works
7 * and redistribute this software and such derivative works
8 * for any purpose, so long as the name of The University of
9 * Michigan is not used in any advertising or publicity
10 * pertaining to the use of distribution of this software
11 * without specific, written prior authorization. If the
12 * above copyright notice or any other identification of the
13 * University of Michigan is included in any copy of any
14 * portion of this software, then the disclaimer below must
17 * THIS SOFTWARE IS PROVIDED AS IS, WITHOUT REPRESENTATION
18 * FROM THE UNIVERSITY OF MICHIGAN AS TO ITS FITNESS FOR ANY
19 * PURPOSE, AND WITHOUT WARRANTY BY THE UNIVERSITY O
20 * MICHIGAN OF ANY KIND, EITHER EXPRESS OR IMPLIED, INCLUDING
21 * WITHOUT LIMITATION THE IMPLIED WARRANTIES OF
22 * MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE. THE
23 * REGENTS OF THE UNIVERSITY OF MICHIGAN SHALL NOT BE LIABLE
24 * FOR ANY DAMAGES, INCLUDING SPECIAL, INDIRECT, INCIDENTAL, OR
25 * CONSEQUENTIAL DAMAGES, WITH RESPECT TO ANY CLAIM ARISING
26 * OUT OF OR IN CONNECTION WITH THE USE OF THE SOFTWARE, EVEN
27 * IF IT HAS BEEN OR IS HEREAFTER ADVISED OF THE POSSIBILITY OF
32 * Portions Copyright (c) 2008
33 * The Linux Box Corporation
36 * Permission is granted to use, copy, create derivative works
37 * and redistribute this software and such derivative works
38 * for any purpose, so long as the name of the Linux Box
39 * Corporation is not used in any advertising or publicity
40 * pertaining to the use or distribution of this software
41 * without specific, written prior authorization. If the
42 * above copyright notice or any other identification of the
43 * Linux Box Corporation is included in any copy of any
44 * portion of this software, then the disclaimer below must
47 * This software is provided as is, without representation
48 * from the Linux Box Corporation as to its fitness for any
49 * purpose, and without warranty by the Linux Box Corporation
50 * of any kind, either express or implied, including
51 * without limitation the implied warranties of
52 * merchantability and fitness for a particular purpose. The
53 * Linux Box Corporation shall not be liable for any damages,
54 * including special, indirect, incidental, or consequential
55 * damages, with respect to any claim arising out of or in
56 * connection with the use of the software, even if it has been
57 * or is hereafter advised of the possibility of such damages.
61 #include <afsconfig.h>
62 #include "afs/param.h"
63 #if defined(AFS_CACHE_BYPASS) || defined(UKERNEL)
64 #include "afs/afs_bypasscache.h"
70 #include "afs/sysincludes.h" /* Standard vendor system headers */
71 #include "afs/afsincludes.h" /* Afs-based standard headers */
72 #include "afs/afs_stats.h" /* statistics */
73 #include "afs/nfsclient.h"
74 #include "rx/rx_globals.h"
77 #define afs_min(A,B) ((A)<(B)) ? (A) : (B)
80 /* conditional GLOCK macros */
81 #define COND_GLOCK(var) \
83 var = ISAFS_GLOCK(); \
88 #define COND_RE_GUNLOCK(var) \
95 /* conditional GUNLOCK macros */
97 #define COND_GUNLOCK(var) \
99 var = ISAFS_GLOCK(); \
104 #define COND_RE_GLOCK(var) \
111 int cache_bypass_strategy = NEVER_BYPASS_CACHE;
112 afs_size_t cache_bypass_threshold = AFS_CACHE_BYPASS_DISABLED; /* file size > threshold triggers bypass */
113 int cache_bypass_prefetch = 1; /* Should we do prefetching ? */
115 extern afs_rwlock_t afs_xcbhash;
118 * This is almost exactly like the PFlush() routine in afs_pioctl.c,
119 * but that routine is static. We are about to change a file from
120 * normal caching to bypass it's caching. Therefore, we want to
121 * free up any cache space in use by the file, and throw out any
122 * existing VM pages for the file. We keep track of the number of
123 * times we go back and forth from caching to bypass.
126 afs_TransitionToBypass(struct vcache *avc,
127 afs_ucred_t *acred, int aflags)
137 if (aflags & TRANSChangeDesiredBit)
139 if (aflags & TRANSSetManualBit)
144 ObtainWriteLock(&avc->lock, 925);
146 * Someone may have beat us to doing the transition - we had no lock
147 * when we checked the flag earlier. No cause to panic, just return.
149 if (avc->cachingStates & FCSBypass)
152 /* If we never cached this, just change state */
153 if (setDesire && (!(avc->cachingStates & FCSBypass))) {
154 avc->cachingStates |= FCSBypass;
158 /* cg2v, try to store any chunks not written 20071204 */
159 if (avc->execsOrWriters > 0) {
160 struct vrequest *treq = NULL;
162 code = afs_CreateReq(&treq, acred);
164 code = afs_StoreAllSegments(avc, treq, AFS_SYNC | AFS_LASTSTORE);
165 afs_DestroyReq(treq);
169 /* also cg2v, don't dequeue the callback */
170 /* next reference will re-stat */
171 afs_StaleVCacheFlags(avc, AFS_STALEVC_NOCB, CDirty);
172 /* now find the disk cache entries */
173 afs_TryToSmush(avc, acred, 1);
174 if (avc->linkData && !(avc->f.states & CCore)) {
175 afs_osi_Free(avc->linkData, strlen(avc->linkData) + 1);
176 avc->linkData = NULL;
179 avc->cachingStates |= FCSBypass; /* Set the bypass flag */
181 avc->cachingStates |= FCSDesireBypass;
183 avc->cachingStates |= FCSManuallySet;
184 avc->cachingTransitions++;
187 ReleaseWriteLock(&avc->lock);
192 * This is almost exactly like the PFlush() routine in afs_pioctl.c,
193 * but that routine is static. We are about to change a file from
194 * bypassing caching to normal caching. Therefore, we want to
195 * throw out any existing VM pages for the file. We keep track of
196 * the number of times we go back and forth from caching to bypass.
199 afs_TransitionToCaching(struct vcache *avc,
209 if (aflags & TRANSChangeDesiredBit)
211 if (aflags & TRANSSetManualBit)
215 ObtainWriteLock(&avc->lock, 926);
217 * Someone may have beat us to doing the transition - we had no lock
218 * when we checked the flag earlier. No cause to panic, just return.
220 if (!(avc->cachingStates & FCSBypass))
223 /* Ok, we actually do need to flush */
224 /* next reference will re-stat cache entry */
225 afs_StaleVCacheFlags(avc, 0, CDirty);
227 /* now find the disk cache entries */
228 afs_TryToSmush(avc, acred, 1);
229 if (avc->linkData && !(avc->f.states & CCore)) {
230 afs_osi_Free(avc->linkData, strlen(avc->linkData) + 1);
231 avc->linkData = NULL;
234 avc->cachingStates &= ~(FCSBypass); /* Reset the bypass flag */
236 avc->cachingStates &= ~(FCSDesireBypass);
238 avc->cachingStates |= FCSManuallySet;
239 avc->cachingTransitions++;
242 ReleaseWriteLock(&avc->lock);
246 /* In the case where there's an error in afs_NoCacheFetchProc or
247 * afs_PrefetchNoCache, all of the pages they've been passed need
251 typedef void * bypass_page_t;
253 #define unlock_and_release_pages(auio)
254 #define release_full_page(pp, pageoff)
257 typedef struct page * bypass_page_t;
259 #define unlock_and_release_pages(auio) \
261 struct iovec *ciov; \
264 afs_int32 iovno = 0; \
265 ciov = auio->uio_iov; \
266 iovmax = auio->uio_iovcnt - 1; \
267 pp = (bypass_page_t) ciov->iov_base; \
270 if (PageLocked(pp)) \
272 put_page(pp); /* decrement refcount */ \
277 ciov = (auio->uio_iov + iovno); \
278 pp = (bypass_page_t) ciov->iov_base; \
282 #define release_full_page(pp, pageoff) \
284 /* this is appropriate when no caller intends to unlock \
285 * and release the page */ \
286 SetPageUptodate(pp); \
290 afs_warn("afs_NoCacheFetchProc: page not locked!\n"); \
291 put_page(pp); /* decrement refcount */ \
296 afs_bypass_copy_page(bypass_page_t pp, int pageoff, struct iovec *rxiov,
297 int iovno, int iovoff, struct uio *auio, int curiov, int partial)
303 dolen = auio->uio_iov[curiov].iov_len - pageoff;
305 dolen = rxiov[iovno].iov_len - iovoff;
307 #if !defined(UKERNEL)
308 # if defined(KMAP_ATOMIC_TAKES_NO_KM_TYPE)
309 address = kmap_atomic(pp);
311 address = kmap_atomic(pp, KM_USER0);
316 memcpy(address + pageoff, (char *)(rxiov[iovno].iov_base) + iovoff, dolen);
317 #if !defined(UKERNEL)
318 # if defined(KMAP_ATOMIC_TAKES_NO_KM_TYPE)
319 kunmap_atomic(address);
321 kunmap_atomic(address, KM_USER0);
326 /* no-cache prefetch routine */
328 afs_NoCacheFetchProc(struct rx_call *acall,
331 afs_int32 release_pages,
336 int moredata, iovno, iovoff, iovmax, result, locked;
345 rxiov = osi_AllocSmallSpace(sizeof(struct iovec) * RX_MAXIOVECS);
346 ciov = auio->uio_iov;
347 pp = (bypass_page_t) ciov->iov_base;
348 iovmax = auio->uio_iovcnt - 1;
349 iovno = iovoff = result = 0;
352 COND_GUNLOCK(locked);
353 code = rx_Read(acall, (char *)&length, sizeof(afs_int32));
354 COND_RE_GLOCK(locked);
355 if (code != sizeof(afs_int32)) {
357 afs_warn("Preread error. code: %d instead of %d\n",
358 code, (int)sizeof(afs_int32));
359 unlock_and_release_pages(auio);
362 length = ntohl(length);
366 afs_warn("Preread error. Got length %d, which is greater than size %d\n",
368 unlock_and_release_pages(auio);
372 /* If we get a 0 length reply, time to cleanup and return */
374 unlock_and_release_pages(auio);
380 * The fetch protocol is extended for the AFS/DFS translator
381 * to allow multiple blocks of data, each with its own length,
382 * to be returned. As long as the top bit is set, there are more
385 * We do not do this for AFS file servers because they sometimes
386 * return large negative numbers as the transfer size.
388 if (avc->f.states & CForeign) {
389 moredata = length & 0x80000000;
390 length &= ~0x80000000;
395 for (curpage = 0; curpage <= iovmax; curpage++) {
397 /* properly, this should track uio_resid, not a fixed page size! */
398 while (pageoff < auio->uio_iov[curpage].iov_len) {
399 /* If no more iovs, issue new read. */
401 COND_GUNLOCK(locked);
402 bytes = rx_Readv(acall, rxiov, &nio, RX_MAXIOVECS, length);
403 COND_RE_GLOCK(locked);
405 afs_warn("afs_NoCacheFetchProc: rx_Read error. Return code was %d\n", bytes);
407 unlock_and_release_pages(auio);
409 } else if (bytes == 0) {
410 /* we failed to read the full length */
412 afs_warn("afs_NoCacheFetchProc: rx_Read returned zero. Aborting.\n");
413 unlock_and_release_pages(auio);
417 auio->uio_resid -= bytes;
420 pp = (bypass_page_t)auio->uio_iov[curpage].iov_base;
421 if (pageoff + (rxiov[iovno].iov_len - iovoff) <= auio->uio_iov[curpage].iov_len) {
422 /* Copy entire (or rest of) current iovec into current page */
424 afs_bypass_copy_page(pp, pageoff, rxiov, iovno, iovoff, auio, curpage, 0);
425 length -= (rxiov[iovno].iov_len - iovoff);
426 pageoff += rxiov[iovno].iov_len - iovoff;
430 /* Copy only what's needed to fill current page */
432 afs_bypass_copy_page(pp, pageoff, rxiov, iovno, iovoff, auio, curpage, 1);
433 length -= (auio->uio_iov[curpage].iov_len - pageoff);
434 iovoff += auio->uio_iov[curpage].iov_len - pageoff;
435 pageoff = auio->uio_iov[curpage].iov_len;
438 /* we filled a page, or this is the last page. conditionally release it */
439 if (pp && ((pageoff == auio->uio_iov[curpage].iov_len &&
440 release_pages) || (length == 0 && iovno >= nio)))
441 release_full_page(pp, pageoff);
443 if (length == 0 && iovno >= nio)
450 osi_FreeSmallSpace(rxiov);
455 /* dispatch a no-cache read request */
457 afs_ReadNoCache(struct vcache *avc,
458 struct nocache_read_request *bparms,
463 struct brequest *breq;
464 struct vrequest *areq = NULL;
468 afs_warn("afs_ReadNoCache VCache Error!\n");
473 /* the receiver will free areq */
474 code = afs_CreateReq(&areq, acred);
476 afs_warn("afs_ReadNoCache afs_CreateReq error!\n");
478 code = afs_VerifyVCache(avc, areq);
480 afs_warn("afs_ReadNoCache Failed to verify VCache!\n");
486 code = afs_CheckCode(code, areq, 11); /* failed to get it */
492 /* and queue this one */
496 breq = afs_BQueue(BOP_FETCH_NOCACHE, avc, B_DONTWAIT, 0, acred, 1, 1,
497 bparms, (void *)0, (void *)0);
502 afs_osi_Wait(10 * bcnt, 0, 0);
514 /* If there's a problem before we queue the request, we need to
515 * do everything that would normally happen when the request was
516 * processed, like unlocking the pages and freeing memory.
518 unlock_and_release_pages(bparms->auio);
520 afs_DestroyReq(areq);
522 osi_Free(bparms->auio->uio_iov,
523 bparms->auio->uio_iovcnt * sizeof(struct iovec));
524 osi_Free(bparms->auio, sizeof(struct uio));
525 osi_Free(bparms, sizeof(struct nocache_read_request));
530 /* Cannot have static linkage--called from BPrefetch (afs_daemons) */
532 afs_PrefetchNoCache(struct vcache *avc,
534 struct nocache_read_request *bparms)
538 struct iovec *iovecp;
540 struct vrequest *areq;
542 struct rx_connection *rxconn;
543 #ifdef AFS_64BIT_CLIENT
544 afs_int32 length_hi, bytes, locked;
548 struct rx_call *tcall;
550 struct AFSVolSync tsync;
551 struct AFSFetchStatus OutStatus;
552 struct AFSCallBack CallBack;
554 struct tlocal1 *tcallspec;
559 iovecp = auio->uio_iov;
562 tcallspec = osi_Alloc(sizeof(struct tlocal1));
564 tc = afs_Conn(&avc->f.fid, areq, SHARED_LOCK /* ignored */, &rxconn);
566 avc->callback = tc->parent->srvr->server;
567 tcall = rx_NewCall(rxconn);
568 #ifdef AFS_64BIT_CLIENT
569 if (!afs_serverHasNo64Bit(tc)) {
570 code = StartRXAFS_FetchData64(tcall,
571 (struct AFSFid *) &avc->f.fid.Fid,
575 COND_GUNLOCK(locked);
576 bytes = rx_Read(tcall, (char *)&length_hi,
578 COND_RE_GLOCK(locked);
580 if (bytes != sizeof(afs_int32)) {
582 COND_GUNLOCK(locked);
583 code = rx_EndCall(tcall, RX_PROTOCOL_ERROR);
584 COND_RE_GLOCK(locked);
588 } /* afs_serverHasNo64Bit */
589 if (code == RXGEN_OPCODE || afs_serverHasNo64Bit(tc)) {
590 if (auio->uio_offset > 0x7FFFFFFF) {
594 pos = auio->uio_offset;
595 COND_GUNLOCK(locked);
597 tcall = rx_NewCall(rxconn);
598 code = StartRXAFS_FetchData(tcall,
599 (struct AFSFid *) &avc->f.fid.Fid,
600 pos, bparms->length);
601 COND_RE_GLOCK(locked);
603 afs_serverSetNo64Bit(tc);
606 code = StartRXAFS_FetchData(tcall,
607 (struct AFSFid *) &avc->f.fid.Fid,
608 auio->uio_offset, bparms->length);
611 code = afs_NoCacheFetchProc(tcall, avc, auio,
612 1 /* release_pages */,
615 afs_warn("BYPASS: StartRXAFS_FetchData failed: %d\n", code);
616 unlock_and_release_pages(auio);
617 afs_PutConn(tc, rxconn, SHARED_LOCK);
621 code = EndRXAFS_FetchData(tcall, &tcallspec->OutStatus,
622 &tcallspec->CallBack,
625 afs_warn("BYPASS: NoCacheFetchProc failed: %d\n", code);
627 code = rx_EndCall(tcall, code);
629 afs_warn("BYPASS: No connection.\n");
631 unlock_and_release_pages(auio);
634 } while (afs_Analyze(tc, rxconn, code, &avc->f.fid, areq,
635 AFS_STATS_FS_RPCIDX_FETCHDATA,
639 * Copy appropriate fields into vcache
643 afs_ProcessFS(avc, &tcallspec->OutStatus, areq);
645 osi_Free(areq, sizeof(struct vrequest));
646 osi_Free(tcallspec, sizeof(struct tlocal1));
647 osi_Free(bparms, sizeof(struct nocache_read_request));
649 /* in UKERNEL, the "pages" are passed in */
650 osi_Free(iovecp, auio->uio_iovcnt * sizeof(struct iovec));
651 osi_Free(auio, sizeof(struct uio));